test(llama): add pinned MiniCPM5-1B coverage

[JiaxinD]

Background

Llama has MiniCPM5-2B coverage but lacks a pinned 1B case. MiniCPM5-1B has hidden width 1536 and attention width 2048; its tokenizer also exposed a prior 18-versus-16-token prefill mismatch.

Exit Criteria

• Select both pinned 1B cases through the existing harness.
• Preserve the HF comparison and raw prompt's 16-token native prefill assertion.
• Build and pass the selected cases on target GPU hardware.

Implementation

• Add FP16 manifests with FP32 references, publisher config fixtures, CPU geometry tests and an explicit release-performance exclusion.
• Add a raw multilingual/digit prompt with fixed HF IDs and a native prefill-count receipt.
• Include the original fix(llama): scan all Split steps for pretokenizer variant classification #1423 tokenizer prerequisite, retaining its author and published history.
• Merge upstream 0fec6d03, preserving speculative/Edge execution and the Qwen embedding exclusion. Keep mode-aware offline snapshot resolution and cache regressions.
• No shared runtime/API/ABI or bundle-format change. The performance exclusion is the only change outside families/llama.

Change categories

• CI or developer tooling
• Model or runtime behavior

Validation

Commands and Results

Command	Result
python -m pytest families/llama/tests/test_checkpoint_cache.py families/llama/tests/test_minicpm5_config.py families/llama/tests/test_runtime_receipt.py qualification_tests/benchmark_qualification/performance/tests -q -p no:cacheprovider --junitxml=minicpm-cpu.xml	570 passed
cmake --build build-sm90 --parallel 8 --target trtmc trtmc_backend_trt trtmc_model_llama test_llama_pretokenizer_variant test_llama_native_kv_cache test_llama_chat_template test_llama_plugin_helpers test_llama_pipeline test_llama_speculative_contract	Passed
ctest --test-dir build-sm90 --output-on-failure -R '^test_llama_(pretokenizer_variant|native_kv_cache|chat_template|plugin_helpers|pipeline|speculative_contract)$'	6 passed, 0 skipped
python -m pytest families/llama/tests/test_e2e.py --e2e-model minicpm5-1b -vv -s --basetemp=/work/e2e-temp --junitxml=/work/evidence/e2e.xml	2 selected cases passed, 8 unrelated cases skipped
python -m tools.model_ci validate, python -m tools.test_impact --validate, git diff --check	Passed

The raw case reports [trtmc.prefill] tokens=16 launches=1 max_chunk=16; all 10 generated IDs match its FP32 HF reference. JUnit, case receipts, source/binary hashes and environment records are retained. No thresholds were changed.

Hardware, Environment, and Revisions

Head d0d47785b92fe881f2a34cb16052396b86e56e32, tree 3c305f19fa5ff6a7b89fc6ef57bb05538649db57. openbmb/MiniCPM5-1B revision 87179e5c1f455ef22e6223592d2d61351b525bfc, Apache-2.0, staged offline. One H200/SM90, Ubuntu 24.04, Python 3.12, driver 595.71.05, CUDA 13.3, TensorRT 11.1.0.106, PyTorch 2.12.0+cu130, Transformers 5.2.0. Image nvcr.io/nvidia/tensorrt:26.07-py3@sha256:b82db1abc23750ab0069abc99bbe4ea29138dbdc23ea39861199e2346638b48a. GPU pinned to the assigned UUID and visible index 0.

Not Run / Remaining Gaps

Long-context, alternate precision, other checkpoints and performance qualification were not run. Current public/protected CI are separate from this local target-hardware result.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

Review #1423 first. The receipt proves prefill count, not every native input ID; the HF generation comparison remains independent. Scoped GPU validation is complete; repository review capacity currently keeps this PR draft.

Risk level

• Medium

Tokenizer behavior and new selected cases affect this checkpoint; broader model/runtime qualification is outside scope.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{Comment @coderabbitai help to get the list of available commands.}